From Claude to In-House Honk: How Spotify is Building an Enterprise AI Coding Collaboration System
At the recent ‘Code w/ Claude’ conference hosted by Anthropic, Niklas Gustavsson, Chief Architect and VP of Engineering at the world’s largest audio streaming platform, Spotify, shared the company’s deep dive into enterprise-level AI coding practices. The core insight of his presentation was that as artificial intelligence significantly accelerates code writing, the real bottleneck in software development will shift from coding itself to areas like codebase governance, validation systems, and organizational collaboration.
New Challenges in the Age of AI Acceleration: From Coding Speed to Engineering Governance
Spotify’s engineering scale is massive, with nearly 3,000 engineers maintaining a large monorepo of about 40 million lines of code and thousands of smaller polyrepos, with approximately 4,500 production deployments daily. AI programming tools have rapidly gained popularity within the company, with over 96% of engineers using them weekly. Among them, 94% believe AI has boosted their personal productivity, directly reflected in an approximate 60% increase in the frequency of pull requests (PRs).
However, the surge in code output has also brought new challenges. In traditional development workflows, coding is often the most time-consuming phase. But with AI assistance, the bottleneck is shifting upstream and downstream: a large volume of PRs needs to be reviewed effectively, code changes must be validated through robust automated systems, and product decisions and release prioritization have become increasingly critical. In short, the comprehensive capabilities of the engineering system are being tested like never before.
The Evolution of Automation: From the Fleetshift Scripting Tool to the Honk Coding Agent
Even before the widespread adoption of AI coding, Spotify had begun to tackle the challenge of large-scale code maintenance. As its codebase was once growing 7 times faster than the number of engineers, a significant amount of engineer time was spent on maintenance tasks such as version upgrades, API deprecations, and security fixes. To address this, Spotify developed the automation tool Fleetshift to execute migration tasks in bulk across multiple code repositories, which has automatically merged about 2.5 million changes to date.
While traditional scripting tools perform well for simple dependency upgrades or configuration changes, they struggle with complex API replacements and changes in invocation patterns. Any externally exposed API might be relied upon by users for any of its observable behaviors (Hyrum’s Law), making the edge cases of large-scale code migrations exceptionally complex.
To solve this problem, Spotify developed its in-house backend coding agent, Honk, based on Anthropic’s Claude model and Agent SDK. Honk runs in Spotify’s own environment, can access trusted internal tools, and can invoke capabilities like CI/CD, build processes, and multi-system validation. By integrating with Fleetshift, Honk can understand and execute complex code modification tasks more intelligently, surpassing the limitations of traditional rule-based scripts.
Towards Collaborative Intelligence: Honk V2 and the Chirp Orchestration Platform
Honk’s application didn’t stop at backend batch tasks. It has gradually evolved into an interactive smart assistant for developers. Engineers can directly call Honk in the internal communication tool Slack, instructing it to complete specific coding tasks and generate a PR. This marks the initial formation of a human-computer interaction model.
Building on this, Spotify launched Honk V2 and integrated it into its internal agent orchestration platform, Chirp. The Chirp platform supports managing multiple parallel agent work sessions and allows team members to share and collaborate on the same AI-assisted task. This model is similar to ‘pair programming’ in software development, but with one party being a human engineer and the other an AI agent that can be co-directed by multiple people, representing an evolution towards more advanced human-AI collaborative intelligence.
The Cornerstone of Standardization: Key to Boosting Both Human and AI Efficiency
One of the most enlightening takeaways from Spotify’s practice is that ‘a standardized codebase’ can simultaneously enhance the efficiency of both human engineers and AI agents.
The company has long been committed to reducing unnecessary technological variations, promoting similar design patterns, technology choices, and project structures for its backend services. For developers, standardization reduces the cognitive load and communication costs of cross-team collaboration. For AI agents, a highly consistent and standardized codebase provides rich ‘local patterns’ from which language models like Claude can learn, enabling them to generate higher-quality code that better conforms to the project’s style and conventions. Practice has shown that Claude performs far better in a structurally consistent codebase than in a fragmented one.
This philosophy is also reflected in the application of its internal developer platform, Backstage. Initially a centralized software catalog that resolved ownership questions like ‘who is responsible for which component,’ Backstage now provides structured information, such as component owners and deployment status, which is accessible to AI agents via APIs or command-line tools. This enables Honk to autonomously query the information it needs and even initiate inquiries with relevant teams via Slack when necessary. This reveals an important trend: a well-designed developer platform can seamlessly evolve into a work platform for AI agents, becoming the underlying infrastructure for human-AI collaboration.
